How to return timestamps?

[cris-m]

I am unable to get timestamps from there result

import whisper
import torch

device = "cuda:0" if torch.cuda.is_available() else "cpu"
model = whisper.load_model("turbo")
audio = whisper.load_audio("TheCubanMissileCrisis.ogg")
audio = whisper.pad_or_trim(audio)
mel = whisper.log_mel_spectrogram(audio, n_mels=128).to(model.device)

# Create a DecodingOptions object
options = whisper.DecodingOptions(task="transcribe", without_timestamps=False)
result = whisper.decode(model, mel, options)

I am getting output like this one:

DecodingResult(audio_features=tensor([[-1.1353e-01,  3.9307e-01, -7.3877e-01,  ...,  9.8228e-04,
         -1.0615e+00,  3.2495e-01],
        [-2.2278e-01, -1.3623e-01, -9.6802e-02,  ...,  4.2633e-02,
          6.6528e-02, -2.6398e-02],
        [-2.4963e-01,  2.0850e-01, -4.0283e-01,  ..., -5.0049e-01,
          4.0552e-01,  2.5562e-01],
        ...,
        [-3.0566e-01, -4.0771e-01,  1.3550e-01,  ..., -4.1675e-01,
          7.1472e-02, -1.6309e-01],
        [-3.0103e-01, -3.7476e-01, -6.2012e-02,  ..., -3.5205e-01,
         -1.5955e-01, -2.0776e-01],
        [-2.6538e-01,  2.0996e-01, -3.7329e-01,  ..., -4.7339e-01,
          3.2007e-01,  1.5002e-01]], device='cuda:0', dtype=torch.float16), language='en', language_probs=None, tokens=[50365, 23956, 264, 2824, 3040, 295, 3374, 6051, 264, 1002, 1768, 295, 11035, 393, 25773, 30515, 40451, 293, 15710, 14909, 322, 264, 644, 295, 604, 4790, 11, 2416, 420, 1359, 13, 50979, 51026, 492, 572, 2854, 1621, 294, 257, 1002, 689, 787, 264, 3539, 16045, 295, 7278, 8855, 257, 11563, 3430, 281, 257, 4790, 311, 3825, 281, 41658, 6674, 46118, 13, 51541, 51590], text="Neither the United States of America nor the world community of nations can tolerate deliberate deception and offensive threats on the part of any nation, large or small. We no longer live in a world where only the actual firing of weapons represents a sufficient challenge to a nation's security to constitute maximum peril.", avg_logprob=-0.12689083814620972, no_speech_prob=1.2446934758922334e-12, temperature=0.0, compression_ratio=1.5330188679245282)

[LaurinmyReha]

This does not directly answer your question but a viable solution to get great timestamps for english especially is this on:

https://github.com/nyrahealth/CrisperWhisper